diff --git a/test/test_tv_tensors.py b/test/test_tv_tensors.py
index ed75ae35ecd..a8e59ab7531 100644
--- a/test/test_tv_tensors.py
+++ b/test/test_tv_tensors.py
@@ -162,6 +162,11 @@ def test_force_subclass_with_metadata(return_type):
     if return_type == "TVTensor":
         assert bbox.format, bbox.canvas_size == (format, canvas_size)
 
+    if torch.cuda.is_available():
+        bbox = bbox.pin_memory()
+        if return_type == "TVTensor":
+            assert bbox.format, bbox.canvas_size == (format, canvas_size)
+
     assert not bbox.requires_grad
     bbox.requires_grad_(True)
     if return_type == "TVTensor":
diff --git a/torchvision/tv_tensors/_torch_function_helpers.py b/torchvision/tv_tensors/_torch_function_helpers.py
index e6ea5fddf35..66812fb5ca6 100644
--- a/torchvision/tv_tensors/_torch_function_helpers.py
+++ b/torchvision/tv_tensors/_torch_function_helpers.py
@@ -69,4 +69,10 @@ def _must_return_subclass():
 
 
 # For those ops we always want to preserve the original subclass instead of returning a pure Tensor
-_FORCE_TORCHFUNCTION_SUBCLASS = {torch.Tensor.clone, torch.Tensor.to, torch.Tensor.detach, torch.Tensor.requires_grad_}
+_FORCE_TORCHFUNCTION_SUBCLASS = {
+    torch.Tensor.clone,
+    torch.Tensor.to,
+    torch.Tensor.detach,
+    torch.Tensor.requires_grad_,
+    torch.Tensor.pin_memory,
+}